OpenAI publishes solutions to more than 370 outstanding math challenges. Math may never be the same

OpenAI says it has published AI-generated full or partial solutions to more than 370 mathematical problems, including some long treated as major open questions. That’s a huge number, and it landed with a thud. Plenty of mathematicians were impressed. Plenty were not, or at least not impressed in the same way.

The company says the work came from an unreleased internal model, and that each solution took about three hours of compute on average. Some of the results are partial, but they include progress on three more Millennium Prize problems. That follows OpenAI’s earlier claim, made weeks ago, that the same kind of internal system had solved the Navier-Stokes equations, one of the seven Clay Mathematics Institute prize problems.

That is exactly why the reaction split so sharply. Supporters see fresh material, new techniques, and a lot of room for human mathematicians to build on. Critics see something uglier: AI companies racing to treat mathematics like a demo reel, while the people who actually work in the field are left to clean up the output and worry about what counts as real progress.

OpenAI tried to head off some of the backlash by publishing the latest results on GitHub and leaning on advice from an independent math advisory group based at the Institute for Advanced Study in Princeton. The group had asked for more disclosure around model names, prompts, reasoning traces, compute time, and why particular problems were chosen. OpenAI shared some of that for 10 problems, plus formalizations for many proofs, but not all of it. The advisory group said the field itself still has to judge whether its recommendations were followed well enough.

There’s a bigger fight underneath all this. Terence Tao says AI is changing math so fast that the field needs a “Math 2.0,” with more emphasis on exposition, community, and new directions rather than just grabbing the next unsolved problem. Dan Litt sounds more upbeat, but even he warns that funding and talent could dry up if people start acting like the discipline has already been automated. In other words: great, the robots can do proofs. Now the humans have to defend the parts of math that still matter.